Critical implementation issues in compensation for nonlinearities in industrial robot manipulators by adaptive multilayer neural networks
Yaolong Lou, Joachim Holtz, Tae-Hee Lee · 1998
To improve the performance of an industrial robot manipulator with linear individual-joint controllers, an adaptive feedforward multilayer neural network (MNN) is proposed as an addition to the existing linear control structure at each joint to compensate the nonlinearity. System stability is guaranteed by three measures: the initialization of the MNN, which ensures that the MNN learning start from a reasonable point; a Lyapunov-based adaptive law in which the MNN is linearized and the residual error is tolerated by a dead-zone or a leakage term; and a contribution function which manipulates the contribution of the MNN to the system. The MNN and the control algorithm are implemented on a TMS320C30 digital signal processor. The realization on a two-link manipulator demonstrates the effectiveness of the proposed scheme.